Artificial intelligence promises to transform the economy. To make that happen, the biggest technology groups are investing hundreds of billions of dollars, putting pressure on their cash flow and seeking increasingly complex sources of financing. The real wager is not whether the technology will prevail. It is whether the returns will justify the cost.
Comparing the present with the past
A technological revolution can change the world and destroy investment capital at the same time. The internet proved it: the technology prevailed, but many of those who financed its first stock market boom never got their money back.
The same question is being asked again today, and the latest earnings reports make it more concrete. Demand for AI is real: Alphabet’s cloud business grew 82% in the second quarter, Microsoft’s Azure 43% and Amazon’s AWS 37%.
At the same time, three of the four biggest players reported negative free cash flow: Alphabet in the second quarter, for the first time since its 2004 stock market listing; Meta in the third quarter, for the first time since 2012; and Amazon on a trailing 12-month basis.
That is the crux: demand may be real and the technology may succeed, but the investments may deliver less than today’s valuations require. The market does not need to question AI. It only needs to question the price it is paying today for tomorrow’s earnings.
1. The infrastructure bill
Big companies keep raising their forecasts for investment in equipment and infrastructure (capex):
| Company | 2026 investment (guidance) | Latest move |
|---|
| Amazon | ~$220bn | Raised from $200bn in the second quarter |
| Alphabet | $195bn–$205bn | Raised in the second quarter |
| Meta | $135bn–$145bn | Lower end raised from $130bn in the third quarter |
| Microsoft | More than $50bn for the current quarter | Guidance for July–September |
The first three alone are guiding to $550bn–$570bn this year. Estimates for all the big companies range from $673bn (UBS, up 76% year on year) to $800bn (FactSet). These figures are forecasts, not final spending.
2. Demand is there. What about cash flow?
The latest reported quarter for each company shows both the strength and the cost of AI in the same table:
| Company (quarter) | Growth | Quarterly investment | Free cash flow |
|---|
| Meta (Q3, 21/10) | Revenue +26%, to $64.6bn | $34.9bn | −$2.1bn |
| Alphabet (Q2) | Google Cloud +82%, to $24.8bn | $44.9bn | −$5.9bn |
| Amazon (Q2) | AWS +37%, to $42.2bn | $54.2bn | −$7.6bn (12 months) |
| Microsoft (FY Q4, to 30/6) | Azure +43% | $35.8bn | +$19.6bn |
The definitions of investment and the time periods covered differ by company.
The figures reveal two truths. The first is positive: AWS grew at its fastest rate in 18 quarters, while Google Cloud achieved an operating margin of 35.6%, with a backlog of $514bn.
The second is the price: Meta invested $34.9bn against operating cash flow of $32.8bn, while Alphabet and Meta have paused share buybacks. Meta’s operating profit was also hit by a $9.9bn legal charge, but the cash flow shortfall stems from investment.
The market is already drawing distinctions. Microsoft, which is funding investment from its cash flow, rose about 15% after its results, while Alphabet fell. Meta lost almost 7% the day after its results, with the cost of insuring against default (CDS) reaching a record high.
The gap is also apparent in forecasts. A Reuters analysis in July, based on LSEG estimates for Microsoft, Alphabet, Amazon, Meta and Oracle, calculated that from 2025 to 2027 additional annual operating cash flow would be about $340bn, against an increase in investment of about $534bn. The market may reward growth for some time, but ultimately it demands a return on capital.
3. From cash to debt
When cash is not enough, borrowing fills the gap. Amazon nearly doubled its long-term debt to $128.9bn from $65.6bn at the end of 2025. Meta ended the third quarter with $83.7bn of debt against $82.4bn in cash. According to IESE, bond issuance by hyperscalers (the cloud giants Amazon, Microsoft, Alphabet, Meta and Oracle) accounts for about 30% of net new investment-grade corporate bond supply in dollars.
The cost of that borrowing matters. On 22 October, Meta’s five-year CDS reached a record high, a sign that creditors are demanding more protection as the company steps up investment. And the financing equation changes as government bond yields rise.
4. Oracle’s example: growth without cash flow
If the giants can afford to cover the gap, Oracle shows what happens when they cannot. Oracle has historically been a software and enterprise database giant. In recent years it has turned into a provider of computing power for AI: it is building data centres and renting them to companies such as OpenAI (the creator of ChatGPT), Elon Musk’s xAI and Meta, competing as a smaller player with the cloud businesses of Amazon, Microsoft and Google.
Its biggest bet is a contract worth about $300bn with OpenAI. In the last financial year, it spent about $55.7bn on data centres and AI infrastructure. Its trailing 12-month free cash flow is negative, at about −$24bn, and it had to issue about $43bn of debt in a year.
The credit market has priced this in. In July, Oracle’s five-year CDS reached about 200 basis points, the highest level since records began in 2008. The company announced a $50bn fundraising programme, including up to $20bn from share sales, while Jefferies does not expect positive free cash flow before fiscal 2029.
Oracle is not simply a borrower. Its growth depends on a customer that itself relies on external financing: Deutsche Bank estimated that OpenAI would burn a cumulative $143bn over 2024–2029. This is the clearest example of the risk described in the headline: strong demand, but capital that is not generated by the business.
5. Private capital and circular deals
Borrowing does not always show up on balance sheets. Meta and Blue Owl, a private credit manager, formed a joint venture for the Hyperion data centre in Louisiana: $27bn of debt in a special-purpose vehicle, with Blue Owl owning 80% and Meta 20%.
The bonds were rated A+, but yielded 6.58% when issued, a rate usually associated with high-yield bonds. IESE estimates that hyperscalers have arranged about $65bn of this kind of off-balance-sheet financing, while Morgan Stanley forecasts $1.15tn in private debt, bonds and securitisations for AI infrastructure over the next four years. For comparison, subprime mortgage securitisations in 2007 amounted to about $2tn.
This does not mean the situation resembles 2007. The borrowers are companies with strong revenues, and the loans are secured by property and contracts. But it does mean the risk is being transferred to insurers, pension funds and private credit — investors who are harder to identify and monitor.
There is also a less visible risk: circular deals. A large company invests in an AI company, which uses the capital to buy computing power from that same company or its partners. These deals make sound economic sense. The risk arises when demand depends on the continuation of the very financing that supports the ecosystem.
In its October 2026 Global Financial Stability Report, the IMF examined 14 such disclosed deals dating from 2023: about $300bn in equity investments, $650bn in debt and $1.4tn in future revenue commitments. These figures do not mean that the amounts are automatically at risk. They do, however, show how concentrated the links are.
8 October showed how fragile this interdependence can be. A Financial Times report said OpenAI had an annualised revenue run rate of about $50bn, against the $70bn the market had expected. One figure from a private company sent the Nasdaq down 1.25% and the semiconductor index down 3.4%, dragging down Nvidia (AI chips), Oracle, CoreWeave (a company that rents out computing power for AI), AMD, Broadcom and Micron. The equal-weighted S&P 500 rose about 0.6% that day.
6. Suppliers: who gets paid first
In the AI supply chain, those selling the tools get paid before those buying them. TSMC’s revenue — the company makes advanced chips for Nvidia and others — is effectively its customers’ investment.
| Link in the chain | What the latest figures showed | The risk |
|---|
| Chips (TSMC) | Third quarter (15/10): revenue of $46.7bn (+41%), net profit +69%; management says demand exceeds supply | Investment of $60bn–$64bn this year; shares fell amid pressure on profit margins |
| Memory (SK Hynix, Micron: memory chips) | Sharp falls on 8 and 13 October | Dependence on customers’ capex and sensitivity to technological shifts, such as the announcement by China’s CXMT |
| Energy (GE Vernova: gas turbines) | Gas turbine orders and reservations of 116 GW in the second quarter, up from 83 GW at the end of 2025; new orders are delivered around 2030–31 | Only about one-fifth is linked to data centres; if demand eases, the order backlog may overstate lasting revenue |
Energy is the most tangible constraint. Morgan Stanley estimates that US data centres will need about 68 GW of new power capacity in 2026–2028, while prices for new gas turbine orders have risen 10–20% since the end of 2025. A data centre without power generates no revenue. And power is not only scarce, it is also more expensive: electricity prices are rising in many parts of the US, increasing the operating costs of every data centre.
Suppliers are the clearest proof that demand exists. But they are also the most exposed link in the chain if their customers revise their investment plans. 8 October showed that.
7. The cost of money
The scale of investment matters more when money becomes more expensive. The US 10-year yield is trading at 5.24%–5.35%, its highest level since 2002, and Brent crude rose above $104 a barrel on 8 October. Higher energy costs fuel inflation and make it harder for interest rates to come down.
Growth companies’ valuations are based on earnings expected far in the future, so they lose value as the discount rate rises, while borrowers seeking to expand pay more. The IMF warns that a decline in AI profitability or investment, combined with leverage and concentrated positions, could trigger a valuation correction, and that a sharp rise in government bond yields is a separate catalyst. This does not mean every interest rate increase leads to a fall. It means there is less room for error.
8. Who can withstand the pressure, and who is exposed
The distinction is not simply between big and small companies. It is between those able to finance their growth and those that depend on continued access to new capital
| Category | Examples | Advantage | Main risk |
|---|
| Large groups with strong cash flow | Microsoft, Alphabet, Amazon | Funding from internal resources, strong cloud growth | Overinvestment, lower returns, suspension of buybacks |
| Companies with high capital needs | Oracle, new infrastructure companies (e.g. CoreWeave) | Direct exposure to demand | Debt, financing costs, dependence on a small number of customers |
| AI model companies | OpenAI | Rapid growth | Dependence on external financing |
| Infrastructure suppliers | TSMC, GE Vernova | Get paid first | Dependence on customers’ investment |
| Businesses using AI | Various | Lower costs, higher productivity | Uncertainty over the scale of actual savings |
The ranking does not predict which companies will succeed. It shows where the risk is concentrated. Even a giant with strong profits can destroy value if it invests too much, while a smaller company can succeed with limited capital if it has a competitive product and recurring revenue. The key is the relationship between the capital required and the value created.
9. What investors should do
Private investors do not need to predict when the AI market will peak. They need to know how dependent their portfolios are on current expectations holding up.
· Check your actual exposure. On 8 October, the Nasdaq fell 1.25%, while the equal-weighted S&P 500 rose about 0.6%. If your index is weighted by market capitalisation, you have more AI risk than the label “diversified” suggests.
· Separate quality from price. An excellent business is not automatically an excellent investment if its price assumes overly optimistic conditions.
· Watch cash flow, not just revenue: investment, net borrowing and return on invested capital.
· Diversify risk factors. Do not abandon AI, but avoid excessive dependence on one group of companies or a single scenario.
· Do not try to predict the date of a correction. Valuation discipline matters more than market timing.
The point is not to stay out of the technology. It is to participate without paying any price and without taking on risk you cannot bear. This article does not constitute investment advice.
10. What will show whether the wager has paid off| What to watch | Positive signal | Warning signal |
|---|
| Free cash flow | Improves as the business scales | Continued deterioration despite revenue growth |
| Cloud growth relative to investment | Revenue accelerating faster than capex | Investment rising faster than economic benefits |
| Financing | Borrowing under control, stable CDS | Growing reliance on debt, private capital and new issuance |
| Quality of demand | Customers paying for measurable value | Demand dependent on financing interconnections |
The next test is close. Meta is so far the only one of the big companies to have reported its third quarter. Results from Alphabet, Microsoft and Amazon are expected in the coming days and will show whether Meta’s picture, with investment exceeding operating cash flow, is an exception or the rule. If cloud growth accelerates enough to cover investment, today’s concerns may prove excessive. If spending continues to rise without a corresponding improvement in cash flow, the risk of valuation revisions will grow.
Conclusion: technology does not guarantee returns
In 10 years, artificial intelligence may be as familiar a part of the economy as electricity and the internet. It will have changed production, services, work and the way businesses operate. But that does not mean everyone financing it today will have made money.
Big groups may create enormous value, but they may also invest more than future returns can justify. Smaller players may develop outstanding technologies and be unable to bear the cost. And investors may be right about the technology but have paid too much for its shares.
That is the difference between technological progress and investment success. The first depends on what the technology can do. The second depends on how much it costs to develop it, who captures the benefits and what price the investor pays.
AI may change the world. The crucial question is who will make money from that change, and who will have paid too much to take part.
Sources: Meta, third-quarter 2026 results and conference call (21/10/2026), and CNBC (22/10/2026) on share price and CDS reaction, Amazon, second-quarter 2026 results (30/7/2026), Alphabet, second-quarter 2026 results (22/7/2026), and Microsoft, fourth-quarter fiscal 2026 results (to 30/6/2026), as covered by 24/7 Wall St and Tom Tunguz, TSMC, third-quarter 2026 results (15/10/2026), International Monetary Fund, Global Financial Stability Report, October 2026, Reuters, 22 July 2026, analysis of free cash flow at big technology groups, Financial Times, 8 October 2026, on OpenAI revenue, as reported by Euronews and other outlets, IESE Business School (30/6/2026) and Columbia Business School / Markus Academy (S. Van Nieuwerburgh, March 2026) on AI infrastructure financing and Morgan Stanley estimates; Dow Jones on the Meta–Blue Owl deal, For Oracle: ICE Data Services (via Thairath), Jefferies (via Globe and Mail), Deutsche Bank (via Motley Fool), GE Vernova, second-quarter 2026 results, and Morgan Stanley on power needs, as reported by oilprice.com and other outlets.
Capital expenditure figures are forecasts, not final results. The amounts for circular financing deals relate to specific disclosed transactions and are not estimates of certain losses.
* Nicholas Havoutis has many years’ experience leading strategic finance functions, having held executive roles at JPMorgan (New York), Chase Manhattan Bank (London) and Eurobank (Athens). He also has a significant presence in the media. Today, as head of SoZone Limited, he advises businesses and investors on international growth, operational optimisation, and mergers and acquisitions strategies.
